In the rapidly advancing field of artificial intelligence (AI), researchers from the University of Cambridge are pioneering a groundbreaking approach to address a fundamental issue – the absence of human uncertainty in AI systems. While AI often excels in precision and accuracy, it falls short when it comes to understanding and incorporating human traits such as doubt and error. In an effort to bridge this gap, scientists are exploring ways to embed human uncertainty into AI programs, particularly in high-stakes domains like medical diagnostics.
Embedding human uncertainty
Many AI systems, especially those that rely on feedback from humans, operate under the assumption that humans are always accurate and certain in their decisions. However, real-life decisions are inherently marked by imperfections, mistakes, and doubts. The research team at the University of Cambridge sought to reconcile this human behavior with machine learning, with the goal of enhancing trustworthiness and reducing risks in AI-human interfaces.
Challenges in handling human uncertainty
To better understand the implications of incorporating human uncertainty into AI systems, the researchers modified an established image classification dataset. They allowed human participants to specify their uncertainty levels while labeling images. The findings were intriguing; while AI systems trained with uncertain labels improved in handling doubtful feedback, the inclusion of human uncertainty sometimes caused a dip in the performance of AI-human systems.
In the realm of AI, the “human-in-the-loop” approach, which allows for human feedback, is seen as a means to mitigate risks in areas where automated systems alone might fall short. However, it raises the question of how these systems respond when human collaborators express doubt. Katherine Collins, the study’s first author, emphasizes the importance of addressing uncertainty from the person’s point of view, stating,
“Uncertainty is central in how humans reason about the world, but many AI models fail to take this into account.”
The study highlights that in scenarios where errors can have minimal consequences, such as mistaking a stranger for a friend, human uncertainty may be inconsequential. However, in safety-sensitive applications, such as clinicians working with medical AI systems, human uncertainty can be perilous. Matthew Barker, a co-author of the study, stresses the need for better tools to recalibrate AI models, empowering individuals to express uncertainty. He notes,
“Although machines can be trained with complete confidence, humans often can’t provide this, and machine learning models struggle with that uncertainty.”
Impact on performance
The research team conducted experiments using various datasets, including one where human participants distinguished bird colors. The results indicated that replacing machine decisions with human ones often led to a significant decline in performance. This demonstrates the challenges in seamlessly integrating humans into machine learning systems.
In the pursuit of bridging the gap between AI and human uncertainty, the researchers are making their datasets publicly available to encourage further exploration in the field. Katherine Collins emphasizes the significance of transparency, stating, “Uncertainty is a form of transparency, and that’s hugely important.” She underlines the need to determine when to trust a model and when to trust a human, especially in applications like chatbots, where understanding the language of possibility is crucial for a natural and safe user experience.
